A new energy power prediction analysis method and system based on model error decoupling

Abstract

This invention discloses a new energy power prediction and analysis method and system based on model error decoupling, comprising: collecting power measurement values, numerical weather forecast data and their version identifiers, measuring point identifiers, seasonal identifiers and measuring point calibration bias information, and generating environmental identifiers to complete traceable grouping of environmental differences such as forecast version upgrades, measuring point replacements and seasonal transitions, thereby suppressing systematic biases caused by environmental aliasing; by correcting power measurement values ​​and constructing update batches according to environmental identifiers to extract invariant feature representations, separating stable correlations from environmental disturbances, and improving cross-environment prediction consistency; by generating a baseline predicted power from the invariant feature representations and constructing a recursive update quantity with the residuals, and cooperating with the recursive least squares estimation of the domain offset compensation quantity, enabling online tracking of domain offset errors; and by synthesizing predicted power and outputting the domain offset intensity index and level and its environmental attribution, reducing scheduling risks and operation and maintenance troubleshooting costs.

Description

Technical Field

[0001] This invention relates to the technical field of power system renewable energy output prediction and meteorological data processing, and in particular to a renewable energy power prediction and analysis method and system based on model error decoupling. Background Technology

[0002] In recent years, with the continuous increase in the penetration rate of new energy sources such as wind power and photovoltaics in the power system, power forecasting has shifted from traditional empirical statistics to a comprehensive technical approach that integrates numerical weather prediction, data assimilation, and machine learning. In engineering applications, the dispatching side typically uses elements such as wind speed, irradiance, temperature, humidity, and cloud cover provided by numerical weather prediction data as external priors, and combines them with historical power measurements from power plants to establish short-term to ultra-short-term forecasting models to support day-ahead / intra-day planning, reserve capacity configuration, and active power control strategies. To improve accuracy and stability, related technologies are gradually incorporating feature engineering, deep learning, and online correction mechanisms. For example, the uncertainty of a single element can be reduced by fusing multiple meteorological elements, the aging of the model can be mitigated by updating the time window, and systematic errors can be repaired by bias correction. However, with the frequent iteration of numerical models by meteorological forecast centers, the normalization of the transformation of station measuring points and the updating of the measurement link, and the changes in meteorological statistical characteristics caused by seasonal transitions, the distribution of predicted inputs exhibits a significant domain shift under different environmental conditions. This has driven power prediction technology to shift from pursuing single-point accuracy to robust prediction and error interpretability analysis oriented towards environmental changes, and has also created a technical demand for the separation of model error sources, continuous updates, and risk classification output.

[0003] The shortcomings of existing technologies are mainly reflected in the following aspects: First, most prediction schemes couple the systematic errors caused by domain offset with errors such as random disturbances and measurement point biases in the same residual, often dealing with them by overall bias correction or retraining. This results in the indivisibility of error sources and the untraceability of correction amounts, making it difficult to form a closed-loop output for prediction, analysis, and operation / schedule decisions. Second, when numerical weather prediction versions are upgraded or resolutions are changed, the statistical distribution, bias structure, and spatiotemporal correlation of meteorological elements will change. Conventional models often map this distribution change as a long-term bias, resulting in cross-seasonal and cross-version accuracy degradation and stability decline, which is more likely to amplify scheduling risks, especially under highly volatile weather conditions. Third, if the measurement bias introduced by measurement point replacement and calibration drift is not explicitly modeled, model updates are likely to absorb the bias as a learnable law into the parameters, causing a lag in adaptation to new measurement points, which in turn leads to an overall rise or fall in the prediction curve. Fourth, many schemes only output predicted power and lack quantitative indicators and level judgments for the intensity of domain offset, making it difficult for the scheduling side to select differentiated backup strategies or trigger targeted model update processes.

[0004] Given that existing new energy power prediction technologies suffer from problems such as difficulty in separating domain offset errors, long-term systematic bias accumulation, and lack of intensity quantification and graded output in scenarios involving numerical weather forecast version upgrades, seasonal changes, and measurement point replacements, this invention is proposed. Therefore, the problem to be solved by this invention is how to construct environmental identifiers and extract invariant feature representations based on the collected power measurements, numerical weather forecast data and their version identifiers, measurement point identifiers, seasonal identifiers, and measurement point calibration bias information. Furthermore, by combining the recursive update amount and the recursive least squares domain offset error estimation, the domain offset compensation amount is obtained, and finally, the predicted power, domain offset intensity index, domain offset level, and prediction analysis results associated with the environmental identifiers are generated. Summary of the Invention

[0005] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section, as well as in the specification abstract and the title of the invention, to avoid obscuring the purpose of this section, the specification abstract, and the title of the invention. Such simplifications or omissions shall not be used to limit the scope of the invention.

[0006] In view of the aforementioned existing problems, the present invention is proposed.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a new energy power prediction and analysis method based on model error decoupling, comprising: collecting power measurement values ​​of new energy power plants, numerical weather forecast data, numerical weather forecast version identifiers, measuring point identifiers, seasonal identifiers and measuring point calibration bias information, and generating environmental identifiers based on the numerical weather forecast version identifiers, the measuring point identifiers and the seasonal identifiers; The power measurement value is corrected according to the measurement point calibration bias information. The numerical weather forecast data and the corrected power measurement value are grouped according to the environmental identifier to construct an update batch. The update batch is then input into the environmental invariant feature extraction model to obtain invariant feature representation. The baseline predicted power of the new energy power station is generated based on the invariant feature representation, and a recursive update quantity is constructed based on the residual between the corrected power measurement value and the baseline predicted power. The recursive update quantity and the environmental identifier are input into the domain offset error estimation model constructed based on the recursive least squares algorithm to obtain the domain offset compensation quantity. A predicted power is generated based on the baseline predicted power and the domain offset compensation amount, and a domain offset intensity index is generated based on the domain offset compensation amount; the domain offset level is determined according to the domain offset intensity index and the preset grading rules, and the prediction analysis results are generated and output in combination with the environmental identifier.

[0008] Secondly, the present invention provides a new energy power prediction and analysis system based on model error decoupling, comprising: The data acquisition and identification generation module is used to acquire power measurement values ​​of new energy power plants through the power acquisition interface, acquire numerical weather forecast data and numerical weather forecast version identifiers corresponding to the numerical weather forecast data through the meteorological data interface, and acquire measuring point identifiers, seasonal identifiers and measuring point calibration bias information, and generate environmental identifiers based on the numerical weather forecast version identifiers, the measuring point identifiers and the seasonal identifiers. The power correction and batch construction module is used to correct the power measurement value according to the measurement point calibration bias information, and to construct and update batches by grouping the numerical weather forecast data and the corrected power measurement value according to the environmental identifier. The invariant feature extraction module is used to perform environment-invariant feature extraction operations on the updated batch to obtain invariant feature representations; The recursive update quantity construction module is used to generate the baseline predicted power of the new energy power station based on the invariant feature representation, and construct the recursive update quantity based on the residual between the corrected power measurement value and the baseline predicted power. The domain offset error estimation module is used to input the recursive update amount and the environmental identifier into the domain offset error estimation model constructed based on the recursive least squares algorithm to obtain the domain offset compensation amount. The predictive analysis output module is used to generate a predicted power based on the benchmark predicted power and the domain offset compensation amount, and generate a domain offset intensity index based on the domain offset compensation amount; determine the domain offset level according to the domain offset intensity index and preset grading rules, generate a predictive analysis result in combination with the environmental identifier, and output the predictive analysis result through the output interface.

[0009] Thirdly, the present invention provides a computer device, comprising: one or more processors; The memory stores operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of the new energy power prediction and analysis method based on model error decoupling as described above.

[0010] Fourthly, the present invention provides a computer-readable medium for storing software, the software including instructions executable by one or more computers, the instructions causing the one or more computers to perform operations, the operations including the flow of the aforementioned new energy power prediction and analysis method based on model error decoupling.

[0011] The beneficial effects of this invention are as follows: By collecting power measurements, numerical weather forecast data and their version identifiers, measuring point identifiers, seasonal identifiers, and measuring point calibration bias information, and generating environmental identifiers, this invention achieves traceable grouping of environmental differences such as forecast version upgrades, measuring point replacements, and seasonal transitions, suppressing systematic biases caused by environmental aliasing; by correcting power measurements and constructing update batches based on environmental identifiers to extract invariant feature representations, stable correlations are separated from environmental disturbances, improving cross-environment prediction consistency; by generating baseline predicted power from invariant feature representations and constructing recursive update quantities using residuals, combined with recursive least squares estimation of domain offset compensation quantities, domain offset errors can be tracked online; by synthesizing predicted power and outputting domain offset intensity indicators and levels and their environmental attributions, scheduling risks and operation and maintenance troubleshooting costs are reduced. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart illustrating the new energy power prediction and analysis method based on model error decoupling as shown in this invention. Detailed Implementation

[0013] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0014] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this invention.

[0015] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0016] According to an embodiment of the present invention, in combination Figure 1 The flowchart shown illustrates a new energy power prediction and analysis method based on model error decoupling, which specifically includes the following steps: S1. Collect power measurement values, numerical weather prediction data, numerical weather prediction version identifiers, measuring point identifiers, seasonal identifiers, and measuring point calibration bias information from the new energy power station, and generate environmental identifiers based on the numerical weather prediction version identifiers, measuring point identifiers, and seasonal identifiers. Note that the following points should be noted in this step: S1.1 At the beginning of each preset sampling period (e.g., 15 minutes), the power measurement value is read through the power acquisition interface (e.g., the station SCADA data interface or the power quality monitoring terminal interface), the numerical weather forecast data and the numerical weather forecast version identifier corresponding to the numerical weather forecast data are read through the meteorological data interface, and the measuring point identifier and measuring point calibration bias information are read from the measuring point configuration table. At the same time, the seasonal identifier is obtained by mapping the date corresponding to the current sampling period according to the seasonal division table.

[0017] As an example, the power measurement value is the average active power over 15 minutes, such as 23.6 MW; the measuring point identifier is the unique identifier in the station measuring point configuration table, such as WF01-PM-001; the measuring point calibration bias information is the combined item of line loss conversion and metering bias corresponding to the measuring point, such as +0.4 MW, which indicates that the reading of the measuring point is lower than the reference metering, and subsequent calibration will be performed additively based on this bias information.

[0018] In a preferred embodiment, the meteorological data interface reads numerical weather forecast data and numerical weather forecast version identifiers corresponding to the sampling period from the forecast data service center; wherein the numerical weather forecast data includes at least meteorological element fields that are highly correlated with new energy power and can be stably obtained, such as: hub height, wind speed, wind direction, air pressure, temperature, air density or relative humidity, and precipitation / cloud cover for wind farms; and total irradiance, direct irradiance, diffuse irradiance, cloud cover, temperature, and wind speed for photovoltaic power stations.

[0019] For example, numerical weather forecast data may include: hub height wind speed 9.8 m / s, wind direction 215 degrees, temperature 6.2 degrees Celsius, air pressure 1008 hPa, and relative humidity 62%. The numerical weather forecast version identifier is the version or batch identifier carried when the forecast service is released, such as 20260122-00Z-V3, which consists of date-start time-version number to distinguish different batches of versions on the same date.

[0020] In a preferred embodiment, the seasonal division table is a pre-set date-to-season mapping table in the system, and seasonal identifiers are generated by month interval → seasonal identifier; for example, March to May is mapped to SPR, June to August to SUM, September to November to AUT, and December to February of the following year to WIN; when the area where the station is located has obvious differences between rainy season and typhoon season, SUM can be further subdivided into SUM-R and SUM-NR, and the mapping result is determined by the date interval table.

[0021] S1.2. Perform coding standardization processing on the logarithmic weather forecast version identifier, measuring point identifier, and seasonal identifier to obtain version code, measuring point code, and seasonal code of uniform length.

[0022] Specifically, encoding standardization includes at least: character set unification, case unification, delimiter unification, length unification, and abnormal character replacement. Among these: The character set and delimiter of the numerical weather forecast version identifier are unified, and it is converted into a string containing only numbers, letters and a single connector; then the length is unified: when the version identifier length is less than the preset length, fixed characters are padded on the left; when it exceeds the preset length, the end is truncated. The same character set and length are uniformly processed for all measurement point identifiers; when the measurement point identifier contains a device type prefix and a serial number segment, it is preferable to retain the device type prefix and serial number segment to improve distinguishability. Perform a fixed enumeration mapping on the seasonal identifiers, mapping the seasonal identifiers to strings of fixed length.

[0023] As an example, the uniform length (i.e., the preset length) in this embodiment is 12 characters: the numerical weather forecast version identifier 20260122-00Z-V3 is normalized to 2026012200V3; the measuring point identifier WF01-PM-001 is normalized to WF01PM0001; and the seasonal identifier WIN is mapped to SEASONWIN000. The above normalization results are all 12 characters long, which facilitates the stable execution of subsequent sequential concatenation and hash mapping.

[0024] S1.3. Concatenate the version code, measurement point code and seasonal code in the order to generate the original value of the environment identifier, and perform hash mapping on the original value of the environment identifier to generate the environment identifier.

[0025] In a preferred embodiment, the original environmental identifier value is concatenated in the order of version code, measurement point code, and season code, and a fixed separator is inserted between adjacent codes to avoid boundary ambiguity; the string obtained after concatenation is used as the input for hash mapping.

[0026] Specifically, the hash mapping in this embodiment uses a hash algorithm with a fixed output length, and the hash result is truncated into a hexadecimal string of a preset length as an environment identifier. To avoid differences in implementation across different systems, the hash mapping also includes a unified encoding method for the input string and a salt value setting method: the input is uniformly encoded in UTF-8; the salt value is a system configuration item and remains fixed within the same deployment.

[0027] For example, the original value of the environment identifier can be 2026012200V3|WF01PM0001|SEASONWIN000; after performing hash mapping, the output environment identifier is EID-7F3A19C2, where EID- is the prefix and 7F3A19C2 is the truncated fixed-length summary fragment; the environment identifier serves as a key field for subsequent grouping, batch updating, and cache retrieval, ensuring that different environment groups can be formed when numerical weather forecast versions, measurement points, or seasons change, thereby avoiding mixing data from different sources into the same recursive state.

[0028] It should be noted that the error in new energy power prediction not only comes from the uncertainty of meteorological input, but also from the systematic shifts caused by forecast version updates, changes in measuring points or metering biases, and seasonal differences in power curves. If the above-mentioned environmental sources are not explicitly distinguished, subsequent model updates will mix the errors from different sources into the same type of error, resulting in recursive state drift and prediction instability. By generating environmental labels at the data entry stage and keeping them stable, the subsequent update process can organize data and maintain the recursive estimation state on an environmental basis, thereby isolating the offset estimates under different environmental conditions and reducing cross-environmental pollution.

[0029] S2. Correct the power measurement values ​​based on the measurement point calibration bias information. Group the numerical weather prediction data and the corrected power measurement values ​​according to environmental identifiers to construct update batches. Input the update batches into the environmental invariant feature extraction model to obtain invariant feature representations. Note that the following should be noted in this step: S2.1 Perform offset correction operation on the power measurement value according to the measurement point calibration offset information to obtain the corrected power measurement value.

[0030] In a preferred embodiment, the measurement point calibration bias information is determined by a combination of the metrological calibration results, the conversion results of the line loss between the grid connection point and the station, and the comparative verification results after the measurement point is replaced. It is stored in the measurement point configuration table and read along with the measurement point identifier. The bias correction operation is performed in the following order: read the bias information → verify with the power measurement value in the same unit → perform additive correction → output the corrected power measurement value.

[0031] As an example, if the power measurement value is 23.6 MW and the measurement point calibration bias information is +0.4 MW, then the corrected power measurement value is 24.0 MW; if the measurement point calibration bias information is -0.3 MW, then the corrected power measurement value is 23.3 MW. When the station uses a piecewise linear calibration table instead of a single bias, the measurement point calibration bias information can be a set of power range-correction entries. During correction, the corresponding correction value is first selected according to the range in which the power measurement value is located, and then additive correction is performed.

[0032] S2.2 Perform element field alignment and dimension consistency processing on the numerical weather forecast data, remove non-meteorological element fields and fill in the missing meteorological element fields according to the filling rules to obtain numerical weather forecast data with consistent meteorological elements.

[0033] As an example, feature field alignment includes: field whitelist selection, field naming mapping, time step alignment, and missing field detection; specifically, the system pre-configures meteorological feature field whitelists, configured separately according to new energy type: Wind power whitelist: hub height wind speed, hub height wind direction, surface air pressure, air temperature, relative humidity, precipitation, boundary layer height; Photovoltaic whitelist: Total irradiance, direct irradiance, diffuse irradiance, cloud cover, temperature, wind speed, and relative humidity.

[0034] Non-meteorological element fields include, but are not limited to, forecast file name, data publisher, download time, interface response code, site description text, and operation and maintenance metadata fields unrelated to meteorological physical quantities; the exclusion rule is that if a field is not on the whitelist and does not belong to the time index field, it will be excluded.

[0035] In a preferred embodiment, the dimensionless processing includes: unit conversion and value range verification; for example, unifying the wind speed unit to meters per second, the temperature unit to degrees Celsius, the air pressure unit to hectopascals, and the irradiance unit to watts per square meter; and marking values ​​that exceed the physically reasonable range according to the outlier rules and filling them in.

[0036] The imputation rules include: prioritizing the use of linear interpolation of adjacent time steps within the same forecast version; when more than 4 consecutive time steps are missing, using the quantile values ​​of historical samples with the same environment identifier to imput them; when the missing field is a critical field and the missing duration exceeds 4 forecast steps, directly marking the sample as an unusable sample and not including it in the update batch.

[0037] S2.3. Based on environmental labels, numerical weather forecast data with consistent meteorological elements and corrected power measurement values ​​are grouped, and samples are selected within each environmental label group according to the preset batch capacity to construct an update batch.

[0038] In a preferred embodiment, each sample record is encapsulated into a sample entry containing an environmental identifier, numerical weather forecast data, corrected power measurement value, and sampling period index; all sample entries are aggregated according to the environmental identifier to obtain multiple environmental groups; within each environmental group, they are sorted in ascending order according to the sampling period index to form a time series.

[0039] Furthermore, the preset batch size is 128 samples. When the number of sample entries in a certain environment group is not less than 128, the most recent 128 samples are selected after sorting by time to construct the update batch. When the number of sample entries is less than 128, it can be handled according to the minimum batch size rule. For example, the minimum batch size is 64. If the number of samples is less than 64, the model parameters will not be updated and only the samples will be cached. An update will be triggered again after the minimum batch size is reached.

[0040] S2.4. Input the updated batch into the environment-invariant feature extraction model to obtain candidate feature representations, calculate the distribution difference consistency term of the candidate feature representations among the updated batches corresponding to different environment identifiers, update the parameters of the environment-invariant feature extraction model based on the distribution difference consistency term and the update error, and output the invariant feature representations.

[0041] In a preferred embodiment, the environment-invariant feature extraction model is constructed with the goal of shared feature extraction backbone + invariance constraint term + prediction error term: the backbone network maps numerical weather forecast data into candidate feature representations; the invariance constraint term measures the consistency of the distribution of candidate feature representations under different environmental labels; the prediction error term measures the error when predicting power from candidate feature representations; this structure enables the model to weaken the dependence on differences in environmental labels while fitting power relationships, and outputs more stable invariant feature representations across different environments.

[0042] For example, the input to the environment-invariant feature extraction model is numerical weather forecast data with consistent meteorological elements, and the output is a candidate feature representation. The model backbone can adopt a two-layer fully connected network or a one-dimensional convolutional network: the first layer performs a nonlinear mapping on each meteorological element to obtain an intermediate representation, and the second layer outputs a candidate feature representation with a fixed dimension; the candidate feature representation has a dimension of 32; its mathematical calculation formula is as follows: in, This is a vector of meteorological elements for a single sample. This forms the backbone of the feature extraction model for environment-invariant feature extraction. Its parameter set; Candidate feature representation; For the overall update goal; This is the prediction error term; These are the invariance constraint weighting coefficients. To update the batch size; For the first The corrected power measurement value corresponding to each sample; The power prediction value is obtained by representing the candidate features. This is the regression head from candidate feature representations to power predictions; For regression head parameters; This is a consistency term for the degree of distributional dissimilarity; A combined index of pairwise environmental identifiers; For environmental labeling The set of candidate feature representations; The mean vector of the candidate feature representation set; It is the L2 norm, used to measure the difference between mean vectors.

[0043] In a preferred embodiment, the process of updating the model parameters based on the distribution difference consistency term and the update error is as follows: Calculate the update batch... Calculation of environmental batches used for comparison Then according to the overall goal right and Perform a parameter update; the update uses an adaptive gradient algorithm; the update error in this embodiment corresponds to The numerical value; for example, when the batch size is 128 and the mean square error is 3.2, it can be used as the prediction error for one update; the output invariant feature is represented as the result of the updated model being forward-computed again, and is written into the feature buffer for subsequent S3 reading within the same sampling period.

[0044] It should be noted that the corrected power measurement value, as the updated label, is closer to the actual grid-connected power, reducing the amplification of errors due to metering bias; grouping by environmental identifier can explicitly isolate the systematic differences brought about by forecast version, measurement point and season, and avoid writing environmental differences into the feature model as an inherent law of meteorological to power mapping.

[0045] Furthermore, by introducing a distribution difference consistency term, the candidate feature representations maintain similar distribution centers under different environments, thereby making the subsequent baseline predicted power more dependent on stable meteorological driving components rather than environmental change components. Compared with existing technologies that only fit a single environment or rely solely on time window updates, this embodiment explicitly constrains cross-environment consistency during the model update stage, which is beneficial for maintaining the stability of feature representations when forecast versions are frequently updated or seasons change.

[0046] S3. Generate the baseline predicted power of the renewable energy power station based on the invariant feature representation, and construct a recursive update quantity based on the residual between the corrected power measurement value and the baseline predicted power; input the recursive update quantity and the environmental label into the domain offset error estimation model constructed based on the recursive least squares algorithm to obtain the domain offset compensation quantity. Note that the following points should be noted in this step: S3.1 Perform a weighted summation operation and a bias correction operation on the invariant feature representation to obtain the baseline predicted power of the new energy power station.

[0047] In a preferred embodiment, the dimension-weighted summation operation is performed in the order of reading the weight table → multiplying dimension by dimension → summing to obtain a scalar, wherein the weight table is a set of parameters calibrated offline, stored in the model configuration area, and associated with the version of the environment-invariant feature extraction model; the bias correction operation is performed in the form of scalar result + bias term, wherein the bias term is used to compensate for the system offset under the method of updating the sample mean.

[0048] For example, the 32-dimensional invariant feature representation is first multiplied dimension by dimension and then summed to obtain 23.2 MW. Then, the bias term +0.6 MW is added to obtain the baseline predicted power of 23.8 MW. The bias term can be configured according to the type of site and the season, such as +0.6 MW for winter and -0.2 MW for summer, to reflect the overall level difference of the baseline curve under different seasons.

[0049] S3.2 Perform a difference calculation on the corrected power measurement value and the benchmark predicted power to obtain the residual, and perform a weighted moving average calculation on the residual within a sliding window consisting of consecutive preset sampling periods to obtain a smoothed residual.

[0050] In a preferred embodiment, the residual is obtained by subtracting the baseline predicted power from the corrected power measurement value; the sliding window is a window length consisting of consecutive preset sampling periods, for example, 16 consecutive sampling periods, i.e., a 4-hour window; the weighted moving average is configured with a weight sequence according to the rule that the more recent the weight, the higher the weight, and the total weight is normalized to 1; when there is a missing measurement residual within the window, the missing measurement item is first removed according to the missing measurement rule, and then the remaining weights are normalized again before the smoothed residual is calculated.

[0051] For example, if the corrected power measurement is 24.0 MW and the baseline predicted power is 23.8 MW, the residual is +0.2 MW. If the smoothed residual calculated within a 4-hour window is +0.35 MW, it indicates that the baseline prediction tends to be underestimated within this window, providing a more stable observation increment for subsequent recursive updates.

[0052] S3.3. Based on the smooth residual and invariant feature representation, perform scaling and pruning operations to generate a recursive update quantity, where scaling uses a preset update step size as a coefficient, and pruning uses a preset upper limit of amplitude to constrain the value range of the recursive update quantity.

[0053] In a preferred embodiment, the recursive update amount is generated jointly by the smoothed residual and the invariant feature representation: first, the smoothed residual is scaled by a preset update step size (e.g., 0.2) to obtain the update candidate value; then, the update candidate value is gated based on the amplitude of the invariant feature representation, and the update candidate value is reduced when the invariant feature representation is not stable enough; finally, the update candidate value is pruned to limit its absolute value to no more than a preset amplitude upper limit (e.g., 2.0 MW).

[0054] For example, when the smoothed residual is +0.35 MW, it is scaled to +0.07 MW; if the gating factor is 0.8, it is +0.056 MW after gating; it is still +0.056 MW after pruning; when an abnormal deviation causes the updated candidate value after gating to be +2.6 MW, the output after pruning is +2.0 MW, so that the recursive estimate will not jump drastically due to a single abnormal residual; thus, the range of the recursive update amount is [-2.0 MW, +2.0 MW].

[0055] S3.4. Based on the environment identifier, retrieve the recursive estimated state set corresponding to the environment identifier in the parameter buffer. The recursive estimated state set includes at least the offset estimate, covariance state matrix, and forgetting factor of the previous sampling period. When the environment identifier is not matched, write the initial offset value, initial covariance matrix, and initial forgetting factor into the parameter buffer and use them as the recursive estimated state set.

[0056] In a preferred embodiment, the parameter cache is organized and stored according to the environment identifier as the primary key, and each primary key corresponds to a recursive estimated state set record; the record includes at least: the offset estimate of the previous sampling period, the covariance state matrix and the forgetting factor, and the most recent update timestamp; the retrieval method is: to perform a hash index lookup in the cache using the environment identifier as the query key; if a match is found, the corresponding state set is read; if no match is found, a new record is created and written to the initial state set.

[0057] For example, the initial offset value is set to 0.0 MW, indicating that no system offset has been observed in the new environment; the initial covariance matrix adopts a diagonal structure, with the main diagonal taking the same initial value, such as 4.0, indicating that the offset estimate is allowed to be adjusted within a large range in the early stage; the initial forgetting factor is set to 0.98, which is used to make recent observations have a greater influence on the estimate.

[0058] S3.5. Using the recursive update amount as the observation increment, the offset estimate in the recursive estimated state set and the covariance state matrix are updated iteratively once according to the recursive least squares update rule to obtain the offset estimate of the current sampling period; and the offset estimate of the current sampling period is determined as the domain offset compensation amount.

[0059] In a preferred embodiment, the process of performing an iterative update according to the recursive least squares update rule includes: (1) Use the recursive update amount as the observation increment input for the current sampling period; (2) Calculate the gain factor based on the covariance state matrix and the forgetting factor, and make the gain factor adaptively adjust with the size of the covariance state matrix; (3) Use the gain factor to correct the offset estimate of the previous sampling period to obtain the offset estimate of the current sampling period; (4) Update the covariance state matrix synchronously, so that it decays over time under the influence of the forgetting factor and is superimposed with the contribution of this observation; (5) Write the updated offset estimate and covariance state matrix back to the corresponding record of the environment identifier in the parameter buffer.

[0060] As an example, the domain offset error estimation model in this embodiment uses the recursive update amount as the observation increment and the regression vector derived from the invariant feature representation as the input term. It performs recursive least squares iterative update within the recursive estimated state set corresponding to the environment identifier. The mathematical expression for calculating the domain offset compensation amount is as follows: in, This indicates the environment index corresponding to the environment identifier; This indicates the sequence number of the current sampling period index within the group corresponding to the environment identifier; Indicates the recursive update quantity; Represents the regression vector. This represents the parameter estimation vector for the current sampling period corresponding to the environmental identifier; This represents the parameter estimation vector for the previous sampling period corresponding to the environmental identifier; This represents the covariance state matrix of the current sampling period corresponding to the environmental identifier; This represents the covariance state matrix of the previous sampling period corresponding to the environmental identifier; This represents the gain vector of the current sampling period corresponding to the environmental identifier; This represents the forgetting factor corresponding to the environmental identifier, with a value range greater than 0 and less than 1. Indicates the transpose operation; This represents the domain offset compensation amount corresponding to the environment identifier, and is used as the domain offset compensation amount output from S3 to S4.

[0061] For example, if the offset estimate of the previous sampling period is +0.30 MW, the recursive update is +0.056 MW, and the current gain factor is equivalent to a correction ratio of 0.6, then the updated offset estimate is +0.33 MW; this offset estimate is then determined as the domain offset compensation amount and used for the S4 synthesized prediction power.

[0062] It should be noted that the baseline predicted power carries the stable and transferable components driven by meteorology, while the domain offset compensation carries the systematic offset components generated by environmental changes. After the two are separated, the impact of environmental changes on the prediction is mainly absorbed by the recursive estimation of the domain offset compensation, thereby avoiding frequent retraining of the main model. Compared with the existing technology that outputs the predicted power end-to-end with only a single model, this embodiment decomposes the error source into a stable part and an offset part, and uses environmental labels to isolate the state in the offset part. This allows for continuous compensation output by simply switching the corresponding cache state when the forecast version is changed or the measurement point is switched, thus improving the continuity of the prediction.

[0063] S4. Generate predicted power based on the baseline predicted power and the domain offset compensation amount, and generate a domain offset intensity index based on the domain offset compensation amount; determine the domain offset level according to the domain offset intensity index and preset classification rules, and generate and output the prediction analysis results in conjunction with the environmental label. Note that the following should be noted in this step: S4.1 Perform a synthesis operation on the reference predicted power and the domain offset compensation amount to obtain the predicted power, and perform amplitude calculation on the domain offset compensation amount to obtain the domain offset intensity index.

[0064] As an example, the formula for calculating the synthesis of predicted power is: in, To predict power, As a baseline predicted power, This is the domain offset compensation amount. For environment indexing, This is the sampling period index.

[0065] As an example, the formula for calculating the magnitude of the domain offset intensity index is: in, For the domain offset intensity index, The result is the magnitude calculation of the absolute value of the domain offset compensation. This is the domain offset compensation amount.

[0066] S4.2. Match the corresponding intensity range in the preset grading rules according to the domain offset intensity index, and determine the domain offset level corresponding to the intensity range.

[0067] Specifically, the preset grading rules include: Obtain the domain offset intensity index set from the historical operation samples corresponding to each environmental identifier. Sort the domain offset intensity index set by numerical value, and take the domain offset intensity index value corresponding to the 70th percentile as the first grade threshold (e.g., 0.45 MW), and take the domain offset intensity index value corresponding to the 90th percentile as the second grade threshold (e.g., 0.90 MW), where the first grade threshold is less than the second grade threshold. Then compare the current domain offset intensity index with the first grade threshold and the second grade threshold. When the domain offset intensity index is less than the first classification threshold, it is determined to be of low level; When the domain offset intensity index is greater than or equal to the first grade threshold and less than the second grade threshold, it is determined to be of medium grade. When the domain offset intensity index is greater than or equal to the second classification threshold, it is determined to be of a high level.

[0068] It should be noted that the sampling window for historical samples can be set to the most recent 90 days, and the sample size should be no less than 2,000. When the sample size is insufficient, the sampling window should be extended until the minimum sample size is met.

[0069] It should also be noted that the low level corresponds to the intensity range from 0 to the first grading threshold; the medium level corresponds to the intensity range from the first grading threshold to the second grading threshold; and the high level corresponds to the intensity range from the second grading threshold and above.

[0070] S4.3 Extract the change source markers corresponding to the environmental markers. The change source markers shall include at least one or more of the following: numerical weather forecast version change, measurement point change, and seasonal change.

[0071] In a preferred embodiment, the change source marker is obtained by comparing the differences between the environmental elements of the current sampling period and the environmental elements of the previous sampling period, specifically including: (1) Read the numerical weather forecast version identifier, measuring point identifier and season identifier of the current sampling period, and read the corresponding field with the same name from the previous sampling period; (2) When the numerical weather forecast version identifier of the current sampling period is inconsistent with that of the previous sampling period, write the numerical weather forecast version change in the change source marker; (3) When the measurement point identifier of the current sampling period is inconsistent with that of the previous sampling period, write the measurement point change in the change source mark; (4) When the seasonal identifier of the current sampling period is inconsistent with that of the previous sampling period, write the seasonal switch in the change source marker; When none of the above three types of differences exist, the change source marker is written as "no explicit change".

[0072] S4.4 Generate the prediction analysis results by combining the predicted power, domain offset intensity index, domain offset level and environmental identifier, and change source marker in that order, and output the prediction analysis results to the display interface.

[0073] It should be noted that by synthesizing the baseline predicted power and the domain offset compensation amount to obtain the predicted power, the predicted output includes both the invariant components driven by meteorology and the recursive compensation components for environmental offsets. Furthermore, by calculating and classifying the domain offset intensity index based on the domain offset compensation amount, and further combining the output with environmental identification and change source markers, the prediction results not only provide numerical values ​​but also the offset level and possible sources, thereby improving the ability of scheduling and operation and maintenance to judge the reliability of the prediction.

[0074] In a preferred embodiment, the output format of the predictive analysis results can be a structured message or a table row, which facilitates unified display on the scheduling end, operation and maintenance end, or monitoring end.

[0075] Furthermore, taking a wind farm WF01 as an example: within the sampling period index of 2026-01-22 10:00, the corrected power measurement value is 24.0 MW, the baseline predicted power is 23.8 MW, the domain offset compensation is +0.33 MW, then the predicted power is 24.13 MW, the domain offset intensity index is 0.33 MW, the environmental identifier is EID-7F3A19C2, therefore the domain offset level is low; if a change in the numerical weather forecast version is detected from 20260122-00Z-V2 to 20260122-00Z-V3, then the change source marker includes a change in the numerical weather forecast version; the final output prediction analysis result is: station WF01, sampling period index 2026-01-22 10:00, predicted power 24.13 MW, domain offset intensity index 0.33 MW, domain offset level low, environmental label EID-7F3A19C2, change source marker numerical weather forecast version change.

[0076] In applying the above embodiments, other aspects of the present invention also propose a new energy power prediction and analysis system based on model error decoupling, comprising: The data acquisition and label generation module is used to acquire power measurement values ​​of new energy power plants through the power acquisition interface, acquire numerical weather forecast data and numerical weather forecast version identifiers corresponding to the numerical weather forecast data through the meteorological data interface, and acquire measuring point identifiers, seasonal identifiers and measuring point calibration bias information, and generate environmental identifiers based on the numerical weather forecast version identifiers, measuring point identifiers and seasonal identifiers. The power correction and batch construction module is used to correct the power measurement values ​​based on the measurement point calibration bias information, and to construct and update batches by grouping numerical weather forecast data and corrected power measurement values ​​according to environmental identifiers. The invariant feature extraction module is used to perform environment-invariant feature extraction operations on the updated batch to obtain invariant feature representations; The recursive update quantity construction module is used to generate the baseline predicted power of new energy power plants based on the invariant feature representation, and construct the recursive update quantity based on the residual between the corrected power measurement value and the baseline predicted power. The domain offset error estimation module is used to input the recursive update amount and the environmental identifier into the domain offset error estimation model constructed based on the recursive least squares algorithm to obtain the domain offset compensation amount. The predictive analysis output module is used to generate predicted power based on the baseline predicted power and the domain offset compensation amount, and to generate a domain offset intensity index based on the domain offset compensation amount; to determine the domain offset level according to the domain offset intensity index and the preset classification rules, to generate predictive analysis results in combination with environmental labels, and to output the predictive analysis results through the output interface.

[0077] Other aspects disclosed in the embodiments of the present invention also provide a computer device including one or more processors and a memory.

[0078] The memory is used to store operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of the new energy power prediction and analysis method based on model error decoupling in the foregoing embodiments, especially... Figure 1 The flowchart of the method is shown.

[0079] Other aspects disclosed in the embodiments of the present invention also propose a computer-readable medium for storing software, the software including instructions executable by one or more computers, the execution of which causes the one or more computers to perform operations, including the flow of the new energy power prediction and analysis method based on model error decoupling of the foregoing embodiments, in particular... Figure 1 The flowchart of the method is shown.

[0080] It should be recognized that embodiments of the present invention may be implemented or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable storage medium.

[0081] The method can be implemented using standard programming techniques, including a non-transitory computer-readable storage medium configured with a computer program in the computer program, wherein the storage medium is configured such that the computer operates in a specific and predefined manner.

[0082] Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system; however, if required, the program can be implemented in assembly or machine language.

[0083] In any case, the language can be either compiled or interpreted.

[0084] Furthermore, for this purpose, the program can run on a programmed application-specific integrated circuit.

[0085] The processes described herein (or variations and / or combinations thereof) can be executed under the control of one or more computer systems configured with executable instructions, and can be implemented by hardware or a combination thereof as code (e.g., executable instructions, one or more computer programs, or one or more applications) that commonly executes on one or more processors. The computer program includes a plurality of instructions executable by one or more processors.

[0086] Furthermore, the method can be implemented in any suitable computing platform, including but not limited to personal computers, minicomputers, mainframes, workstations, networked or distributed computing environments, standalone or integrated computer platforms, or in communication with charged particle tools or other imaging devices.

[0087] Various aspects of the present invention can be implemented in machine-readable code stored on a non-transitory storage medium or device, whether portable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it can be read by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the processes described herein.

[0088] Furthermore, machine-readable code, or parts thereof, can be transmitted via wired or wireless networks.

[0089] When such media includes instructions or programs that combine with a microprocessor or other data processor to implement the steps described above, the invention described herein includes these and other different types of non-transitory computer-readable storage media.

[0090] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A new energy power prediction and analysis method based on model error decoupling, characterized in that, include: Collect power measurement values, numerical weather forecast data, numerical weather forecast version identifiers, measuring point identifiers, seasonal identifiers, and measuring point calibration bias information of new energy power stations, and generate environmental identifiers based on the numerical weather forecast version identifiers, the measuring point identifiers, and the seasonal identifiers; The power measurement value is corrected according to the measurement point calibration bias information. The numerical weather forecast data and the corrected power measurement value are grouped according to the environmental identifier to construct an update batch. The update batch is then input into the environmental invariant feature extraction model to obtain invariant feature representation. The baseline predicted power of the new energy power station is generated based on the invariant feature representation, and a recursive update quantity is constructed based on the residual between the corrected power measurement value and the baseline predicted power. The recursive update quantity and the environmental identifier are input into the domain offset error estimation model constructed based on the recursive least squares algorithm to obtain the domain offset compensation quantity. A predicted power is generated based on the baseline predicted power and the domain offset compensation amount, and a domain offset intensity index is generated based on the domain offset compensation amount; the domain offset level is determined according to the domain offset intensity index and the preset grading rules, and the prediction analysis results are generated and output in combination with the environmental identifier.

2. The new energy power prediction and analysis method based on model error decoupling according to claim 1, characterized in that, The method for generating the environmental identifier includes: At the beginning of each preset sampling period, the power measurement value is read through the power acquisition interface, the numerical weather forecast data and the numerical weather forecast version identifier corresponding to the numerical weather forecast data are read through the meteorological data interface, and the measuring point identifier and the measuring point calibration bias information are read from the measuring point configuration table. At the same time, the seasonal identifier is obtained by mapping the date corresponding to the current sampling period according to the seasonal division table. The numerical weather forecast version identifier, the measuring point identifier, and the seasonal identifier are respectively subjected to coding standardization processing to obtain version code, measuring point code, and seasonal code of uniform length; The environment identifier is generated by concatenating the version code, the measurement point code, and the season code in that order, and then performing a hash mapping on the environment identifier to generate the environment identifier.

3. The new energy power prediction and analysis method based on model error decoupling according to claim 1, characterized in that, The method for generating the invariant feature representation includes: Based on the measurement point calibration bias information, a bias correction operation is performed on the power measurement value to obtain the corrected power measurement value; The numerical weather forecast data is subjected to element field alignment and dimension consistency processing. Non-meteorological element fields are removed and missing meteorological element fields are filled in according to the filling rules to obtain the numerical weather forecast data with consistent meteorological elements. Based on the environmental identifier, the numerical weather forecast data with consistent meteorological elements and the corrected power measurement value are grouped, and samples are selected in each group corresponding to the environmental identifier to construct an update batch according to a preset batch capacity. The updated batch is input into the environment-invariant feature extraction model to obtain candidate feature representations. The distribution difference consistency term of the candidate feature representations is calculated among the updated batches corresponding to different environment identifiers. The parameters of the environment-invariant feature extraction model are updated according to the distribution difference consistency term and the update error, and the invariant feature representation is output.

4. The new energy power prediction and analysis method based on model error decoupling according to claim 3, characterized in that, The method for constructing the recursive update quantity is as follows: The invariant feature representation is subjected to a dimension-weighted summation operation and a bias correction operation to obtain the baseline predicted power of the new energy power station; The difference between the corrected power measurement value and the benchmark predicted power is calculated to obtain the residual, and a weighted moving average is calculated on the residual within a sliding window consisting of consecutive preset sampling periods to obtain a smoothed residual. Based on the smoothed residual and the invariant feature representation, scaling and pruning operations are performed to generate a recursive update quantity, wherein the scaling uses a preset update step size as a coefficient, and the pruning constrains the value range of the recursive update quantity with a preset upper limit of magnitude.

5. The new energy power prediction and analysis method based on model error decoupling according to claim 4, characterized in that, The method for generating the domain offset compensation includes: Based on the environment identifier, a recursive estimation state set corresponding to the environment identifier is retrieved from the parameter cache. The recursive estimation state set includes at least the offset estimate, covariance state matrix, and forgetting factor of the previous sampling period. When the environment identifier is not matched, the initial offset value, initial covariance matrix, and initial forgetting factor are written into the parameter cache and used as the recursive estimation state set. Using the recursive update amount as the observation increment, the offset estimate in the recursive estimated state set and the covariance state matrix are iteratively updated once according to the recursive least squares update rule to obtain the offset estimate of the current sampling period; and the offset estimate of the current sampling period is determined as the domain offset compensation amount.

6. The new energy power prediction and analysis method based on model error decoupling according to claim 5, characterized in that, The method for generating the predictive analysis results includes: The predicted power is obtained by performing a synthesis operation on the reference predicted power and the domain offset compensation amount, and the domain offset intensity index is obtained by performing amplitude calculation on the domain offset compensation amount. Based on the domain offset intensity index, match the corresponding intensity interval in the preset grading rules to determine the domain offset level corresponding to the intensity interval; Based on the environmental identifier, the change source marker corresponding to the environmental identifier is extracted. The change source marker includes at least one or more of the following: numerical weather forecast version change, measurement point change, and seasonal change. The predictive analysis results are generated by combining the predicted power, the domain offset intensity index, the domain offset level, the environmental identifier, and the change source marker in that order, and then output to the display interface.

7. The new energy power prediction and analysis method based on model error decoupling according to claim 6, characterized in that, The preset grading rules include: A set of domain offset intensity indicators is obtained from the historical operation samples corresponding to each of the aforementioned environmental identifiers. The set of domain offset intensity indicators is sorted by numerical value, and the domain offset intensity indicator value corresponding to the 70th percentile is taken as the first grading threshold, and the domain offset indicator value corresponding to the 90th percentile is taken as the second grading threshold, wherein the first grading threshold is less than the second grading threshold. The current domain offset intensity indicator is then compared with the first grading threshold and the second grading threshold. When the domain offset intensity index is less than the first classification threshold, it is determined to be of a low level; When the domain offset intensity index is greater than or equal to the first grading threshold and less than the second grading threshold, it is determined to be of medium level; When the domain offset intensity index is greater than or equal to the second grading threshold, it is determined to be of a high level.

8. A new energy power prediction and analysis system based on model error decoupling, applied to the new energy power prediction and analysis method based on model error decoupling as described in any one of claims 1 to 7, characterized in that, include: The data acquisition and identification generation module is used to acquire power measurement values ​​of new energy power plants through the power acquisition interface, acquire numerical weather forecast data and numerical weather forecast version identifiers corresponding to the numerical weather forecast data through the meteorological data interface, and acquire measuring point identifiers, seasonal identifiers and measuring point calibration bias information, and generate environmental identifiers based on the numerical weather forecast version identifiers, the measuring point identifiers and the seasonal identifiers. The power correction and batch construction module is used to correct the power measurement value according to the measurement point calibration bias information, and to construct and update batches by grouping the numerical weather forecast data and the corrected power measurement value according to the environmental identifier. The invariant feature extraction module is used to perform environment-invariant feature extraction operations on the updated batch to obtain invariant feature representations; The recursive update quantity construction module is used to generate the baseline predicted power of the new energy power station based on the invariant feature representation, and construct the recursive update quantity based on the residual between the corrected power measurement value and the baseline predicted power. The domain offset error estimation module is used to input the recursive update amount and the environmental identifier into the domain offset error estimation model constructed based on the recursive least squares algorithm to obtain the domain offset compensation amount. The predictive analysis output module is used to generate a predicted power based on the benchmark predicted power and the domain offset compensation amount, and generate a domain offset intensity index based on the domain offset compensation amount; determine the domain offset level according to the domain offset intensity index and preset grading rules, generate a predictive analysis result in combination with the environmental identifier, and output the predictive analysis result through the output interface.

9. A computer device, characterized in that, include: One or more processors; The memory stores operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of the new energy power prediction and analysis method based on model error decoupling as described in any one of claims 1 to 7.

10. A computer-readable medium for storing software, characterized in that: The software includes instructions executable by one or more computers, which cause the one or more computers to perform operations, including the flow of the new energy power prediction and analysis method based on model error decoupling as described in any one of claims 1 to 7.

Images